A Market That Refuses to Cool
Carmen Li left Bloomberg with a specific objective. Track GPU pricing in a market where demand has outpaced every expectation.
Her company, Silicon Data, now measures shifts across AI infrastructure, and the latest data points in one direction. Prices are not stabilizing. They are climbing.
Rising Prices Across Key Chips
Data from Silicon Data shows consistent increases across major GPU categories from Nvidia.
The Neo Cloud H100 index rose from 2.20 to 2.64 in three months, marking a 20 percent increase. The B200 index climbed 22 percent, while hyperscaler H100 pricing increased by 3 percent.
These movements reflect a broader imbalance. Demand for compute continues to exceed available capacity.
No Normal Pricing Cycle
In previous hardware cycles, prices followed a predictable pattern. New chips launched at high prices, then declined as supply expanded.
That pattern has not emerged here.
Even as newer GPUs enter the market, pricing for both new and older chips remains elevated. The H100, despite not being the newest generation, continues to command strong demand for both training and inference workloads.
Pressure Across the Entire Stack
The constraint is not limited to GPUs. It extends across the entire AI infrastructure layer.
Memory supply, power availability, and data center capacity all contribute to the bottleneck. Each constraint reinforces the others, preventing any meaningful easing of prices.
This interconnected pressure sustains elevated costs across the system.
Cloud Providers Command Premium Pricing
The highest pricing appears within hyperscaler environments. Companies such as Amazon, Microsoft, Google, and Oracle charge significantly more for GPU rentals.
Customers pay for integrated infrastructure, reliability, and immediate access to capacity.
In comparison, specialized providers such as CoreWeave offer lower-cost alternatives, though often with narrower service layers.
Minimal Depreciation Signals Strong Demand
One of the clearest indicators of sustained demand lies in resale values.
According to Li, a refurbished H100 retains around 85 percent of its value in the second year and 84 percent in the third year.
This level of retention is unusual for hardware. It signals that demand remains strong enough to absorb supply even in secondary markets.
Implications for the AI Economy
Rising GPU prices affect more than infrastructure budgets. They influence the cost structure of AI products themselves.
Companies relying on rented compute must account for higher operating costs when generating outputs at scale.
At the same time, firms investing heavily in data centers benefit from stable or rising asset values, reducing the risk of rapid depreciation.
Direction of the Market
The current trajectory reflects a single condition. Demand for AI compute continues to expand faster than supply can respond.
Until capacity catches up across chips, power, and infrastructure, pricing pressure is likely to persist.
Carmen Li, CEO of Silicon Data. Carmen Li
Source: BI
